AI worm debate: Saxe argues agents could hide in $100B inference stream, Halvar pushes back on cost and capability limits
joshua_saxe · x · 2026-09-14
Security researcher Joshua Saxe rebuts Halvar Flake's skepticism about AI worms: the worm wouldn't need to generate novel exploits at first, and could spread via social engineering, known exploits, and credential stuffing before accumulating resources. He notes 100bn+ dollars are spent on inference compute today, so malicious agents could plausibly hide in that traffic. Halvar counters that stolen keys would quickly hit cost limits and anomaly detection, good 0days cost thousands of dollars, and current models are empirically bad at administering even small machine fleets.
Related event: Saxe Defends Feasibility of Self-Replicating AI Botnets(3 posts)→
More from AGI Musings
- dhh on agency: 'It's rarely given. It's taken, then recognized' — jamesbrooksco · 2026-09-14
- 'I feel less alone': AI safety anxiety becomes a shared community mood — ShakeelHashim · 2026-09-14
- Investor slams Anthropic's doom discourse as 'sophisticated grift' — StewartalsopIII · 2026-09-14
- Musk says he contributed to Bostrom's Superintelligence, cited by name in foreword — elonmusk · 2026-09-14
- Toby Walsh's week of AI chaos: hijacked sites, Hugging Face hack, bioweapon requests — TobyWalsh · 2026-09-14
- Musk endorses warning: most people in the real world have no idea what's coming — elonmusk · 2026-09-14